Implementation guide
OWASP AISVS v1.01
Adversarial Robustness · Membership-Inference and Model-Inversion Mitigation
Calibrate the model's confidence
Verify that model outputs are calibrated to reduce overconfident predictions.
OWASP AISVS v1.01, requirement 11.2.3
In plain English
Calibrate the confidence scores your model reports, so that "99% sure" is right about 99% of the time. Overconfident models make privacy attacks easier, because records from the training data stand out by how certain the model is about them.
What good looks like
- Calibration is measured on held-out data, for example as expected calibration error.
- Where it is poor, outputs are adjusted.
- Calibration is rechecked after retraining or fine-tuning.
- Scores shown outside the system are the calibrated ones, and only as precise as needed.
How teams usually do it
- Temperature scaling, a simple adjustment applied after training, or Platt scaling or isotonic regression for classifiers.
- Label smoothing or regularisation during training, which reduces overconfidence at source.
- Rounding or bucketing scores before returning them.
Suggested evidence
- Calibration measurements before and after.
- The calibration method and its parameters.
Common mistakes
- Exposing raw probability scores the model was never calibrated for.
- Calibrating once and never rechecking after a retrain.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.